AI Email Triage: From 200 Unread to 5 Decisions

Email triage is the most requested agent we build, and the most misunderstood. Clients ask for "AI that answers my email." What they actually need is AI that makes 195 of 200 emails disappear from their attention and turns the remaining 5 into prepared decisions. Answering is the easy part; deciding what deserves an answer is the product.

The Triage Tiers

Every inbound email gets classified into one of five tiers. The tier determines what happens next — and crucially, how much human attention it consumes.

The "unsure" tier is not a failure state — it's the safety valve. We set the confidence threshold deliberately conservative at launch (escalate below 85%) and lower it as the correction data accumulates.

Context Is What Makes Drafts Usable

A draft written from the email alone reads like a stranger wrote it. A draft written with context reads like you. Before drafting, our agents pull:

• The full thread history, not just the last message
• The sender's entity record: client status, open matters, recent interactions across all connected systems
• Related open tasks and documents ("the contract they're asking about is out for signature since Tuesday")
• Your last 20 replies to similar emails, for tone and typical commitments

The difference is visible in edit rates. Context-free drafts get rewritten ~60% of the time. Context-rich drafts in our production systems get sent with zero or trivial edits roughly 4 times out of 5.

The Approval Queue Is the Product Surface

Triage output lands in a single queue, not scattered across the inbox. Each card shows the classification, the reasoning with a confidence percentage, and the proposed action. The user's job shrinks to: approve, edit, reject, or reclassify. Every one of those verbs is a training signal:

1. Approve — reinforces the classification and the draft style
2. Edit — diffs feed the voice model; recurring edits become standing rules
3. Reject — the agent asks one clarifying question, then updates rules
4. Reclassify — sender or topic rules update immediately, not at the next retrain

Metrics That Tell You It Works

• Escalation rate: percentage of emails needing any human attention. Healthy trajectory: ~40% in week one, under 15% by month two.
• Draft acceptance rate: sent-with-minor-edits or better. Target 75%+.
• Missed-important rate: emails the human dug out of the archive. This is the trust metric — it must be near zero, and every miss deserves a post-mortem.
• Time-to-inbox-zero: the number users actually feel. We routinely see 90 minutes a day drop to under 15.

Where to Start

Don't start with drafting — start with classification in shadow mode. Run the classifier for a week, show the tiers it would have assigned, and count the disagreements. It builds trust, it tunes the thresholds on real data, and it costs nothing when it's wrong. Drafting comes second, sending third, and auto-sending only for tiers with months of clean approvals. Triage done right isn't about AI writing your email. It's about you making 5 decisions instead of opening 200 messages.